Breast tumor classification method and device based on ultrasonic image two-stage deep learning

Abstract

The invention discloses a breast tumor classification method and device based on ultrasonic image two-stage deep learning, and the method comprises the steps: obtaining a BI-RADS tag image set D1 from a breast tumor ultrasonic image, and obtaining a biopsy tag image set D2 from the BI-RADS tag image set D1; training the deep neural network M1 and the deep neural network M2 through the BI-RADS tag image set D1 and the biopsy tag image set D2; using the trained deep neural network M1 parameters to initialize the parameters of the deep neural network M2 with the same network structure as the deep neural network M1; classifying the BI-RADS labels of the breast tumor ultrasonic images to be classified; and when the BI-RADS tag is at level 3 or above, classifying the biopsy tag of the breast tumor ultrasonic image to be classified, and determining whether the breast tumor ultrasonic image to be classified is benign or malignant. The method and device can effectively improve the accuracy and intelligent level of breast tumor ultrasonic image classification diagnosis, and can be used in the fields of assisting ultrasonic doctors in medical diagnosis and the like.

Description

technical field [0001] The present invention relates to the technical field of computer-aided breast ultrasound image tumor automatic diagnosis, and more specifically relates to a breast tumor classification method and device for ultrasound image two-stage deep learning. Background technique [0002] Traditional image-based computer-aided diagnosis systems generally define the image features of breast tumors manually, and then classify them as benign or malignant. Recently, deep learning has been used to diagnose ultrasound images of breast tumors. It has two main advantages: First, deep learning can directly discover features from ultrasound images, making up for the knowledge limitations of manually defined features. Second, deep learning is an end-to-end way to learn image features, with a higher degree of automation and better performance. [0003] The application of deep learning methods requires a learning dataset of breast ultrasound images containing tumor classific...

AI Technical Summary

Problems solved by technology

Shen et al. proposed a sample selection strategy, that is, in each training iteration, delete the data part with the largest loss, and update the model parameters to minimize the loss function on the remaining training data. This method assumes that the model gradually converges, and the training A classifier with good knowledge discrimination ability is produced, so as the training progresses, the training samples with wrong labels will show a higher loss value, but the sample selection scheme usually encounters the problem of sample selection bias, which leads to the trained network Instead, learn the wrong knowledge

[0005] The existence of noise labels will have an irreversible negative impact on the performance of the system. Although the sample weighting method and sample selection method can reduce the influence of noise labels to a certain extent in the general image field, in breast tumor ultrasound image sets, accurate label data The amount is much smaller than the amount of labeled data containing noise

Therefore, it is difficult to design a reasonable sample weighting scheme for breast tumor ultrasound image training data, and the sample selection scheme usually encounters the problem of sample selection bias, which will lead to a decrease in system performance

Images